Open Access Research

Three papers. Peer-reviewed foundations.

SuperLocalMemory is the only agent memory system with published mathematical proofs. Every claim is citable. Every algorithm is verifiable.

3
Papers
2603–2604
arXiv series
2026
Latest publication
01
arXiv:2603.14588 2026 March 2026

SuperLocalMemory V3: Information-Geometric Agent Memory with Adaptive Lifecycle Management

Varun Pratap Bhardwaj

This paper introduces information-geometric principles to agent memory management, replacing cosine-similarity heuristics with the Fisher-Rao distance metric on the statistical manifold. We demonstrate that confidence-weighted geodesic distances improve retrieval precision by 23% over baseline systems. The adaptive lifecycle manager uses Riemannian parallel transport to consolidate coherent memories and decay isolated facts without arbitrary TTL parameters.

Key Contributions
  • Fisher-Rao retrieval metric replacing cosine similarity
  • Riemannian manifold lifecycle (no arbitrary TTL)
  • 74.8% LoCoMo score (local-first, Mode A)
  • 87.7% LoCoMo score (full power, Mode C)
Full Paper
Read on arXiv 
@article{bhardwaj2026slmv3,
  title={SuperLocalMemory V3: Information-Geometric
    Agent Memory with Adaptive Lifecycle Management},
  author={Bhardwaj, Varun Pratap},
  journal={arXiv preprint arXiv:2603.14588},
  year={2026}
}
02
arXiv:2603.02240 2026 March 2026

SuperLocalMemory V2: Bounded Persistent Memory for Autonomous Agents with Tiered Storage Architecture

Varun Pratap Bhardwaj

We present a tiered storage architecture for agent memory that maintains O(log n) retrieval performance regardless of memory store size. The system partitions memories across hot (graph), warm (vector), and cold (compressed) tiers using statistical access patterns. Empirical evaluation over five years of continuous deployment demonstrates zero performance degradation at 1M+ fact scales.

Key Contributions
  • CozoDB graph store for relationship-rich hot tier
  • LanceDB vector index for warm semantic retrieval
  • 32x cold compression with byte-exact reversibility
  • 1M+ facts, O(log n) retrieval, zero slowdown
Full Paper
Read on arXiv 
@article{bhardwaj2026slmv2,
  title={SuperLocalMemory V2: Bounded Persistent Memory
    for Autonomous Agents with Tiered Storage Architecture},
  author={Bhardwaj, Varun Pratap},
  journal={arXiv preprint arXiv:2603.02240},
  year={2026}
}
03
arXiv:2604.06392 2026 April 2026

Qualixar OS: An Agent Operating System Architecture for AI Reliability Engineering

Varun Pratap Bhardwaj

Qualixar OS defines a reference architecture for composable AI reliability tooling, treating agent memory, behavioral assertions, evaluation, and skill management as first-class OS primitives. We introduce the concept of AI Reliability Engineering as a discipline distinct from ML engineering, and demonstrate the architecture with SuperLocalMemory as the memory primitive.

Key Contributions
  • AI Reliability Engineering as a formal discipline
  • Agent OS architecture with composable reliability primitives
  • Information-theoretic prompt compression bounds
  • Reference implementation: SuperLocalMemory V3
Full Paper
Read on arXiv 
@article{bhardwaj2026qualixaros,
  title={Qualixar OS: An Agent Operating System Architecture
    for AI Reliability Engineering},
  author={Bhardwaj, Varun Pratap},
  journal={arXiv preprint arXiv:2604.06392},
  year={2026}
}
2603.14588 — V3 SLM: Information-Geometric Agent Memory 2603.02240 — V2 SLM: Bounded Persistent Memory 2604.06392 — Qualixar OS Architecture
BibTeX

Cite this work

All three papers are open access on arXiv. The primary citation for SuperLocalMemory V3 is below.

arXiv:2603.14588 — Primary Reference 
@article{bhardwaj2026slmv3,
  title={SuperLocalMemory V3: Information-Geometric Agent Memory
    with Adaptive Lifecycle Management},
  author={Bhardwaj, Varun Pratap},
  journal={arXiv preprint arXiv:2603.14588},
  year={2026}
}
Methodology

Why mathematics?

Most agent memory systems are built on heuristics. SLM is built on proofs. Here is why that distinction matters.

01 · The problem with cosine

Cosine similarity ignores confidence.

Cosine similarity treats all embeddings equally regardless of the certainty behind them. A high-confidence fact and a speculative guess are indistinguishable. In agent contexts where memory compounds across sessions, this degrades over time — the geometry is wrong for the task.

02 · Fisher-Rao respects probability

The manifold knows what cosine doesn't.

The Fisher-Rao metric is the natural distance on the statistical manifold — the space where probability distributions live. Confidence scores become weights in a geodesic distance calculation. High-certainty memories are geometrically closer to where they should be. The metric is provably invariant to reparametrization.

See the full mathematical proof →
03 · Peer review matters

Citable claims, not marketing.

Every performance figure in this site — 74.8% LoCoMo, 87.7% full-power, 23% precision improvement over baseline — is cited to a specific arXiv paper. Peer review ensures the methodology is sound and the results are reproducible by any researcher. This is AI Reliability Engineering, not AI marketing.

Read the proofs →

All research is open access. All code is AGPL v3.

Every paper is freely available on arXiv. The full implementation is published on GitHub under the GNU Affero General Public License v3. No paywalls, no proprietary weights, no hidden state. Reproducibility is the point.

GitHub All Papers on arXiv